生成式人工智能训练数据著作权“合理使用”的边界认定与规则调适

Delimiting the Boundary of Copyright “Fair Use” for Generative AI Training Data and Adapting the Rules

摘要: 生成式人工智能的模型训练需要大规模利用已有作品，此类利用能否构成合理使用，直接决定训练数据的合规成本与权利人的受偿范围。既有研究与司法裁判分别以使用阶段、行为性质和主体目的划定边界，三种划法结论互斥而标准无法通约，根本原因在于划界依据取自产业政策权衡，而非著作权的权利构造本身。著作权的排他效力仅及于作品的表达性利益，边界的划定因而应以使用行为是否形成表达性替代为准。具体来说，输入与训练阶段的非表达性利用不进入著作权权利范围，输出端构成实质性相似的表达再现落入专有权范围，模型记忆的评价位置在输出端而非训练端，训练语料许可市场上的收益预期因循环论证与定价对象的缺失，不属于三步检验法所称的合法权益。受制于复制权的过度扩张与《著作权法》第24条的封闭式列举，该判准在现行法上尚无适用空间。我国宜将作品性使用规定为著作权侵权的成立要件，并在《著作权法实施条例》中增设附机器可读权利保留机制的文本与数据挖掘例外。

Abstract: Training generative AI models requires the large-scale use of existing works. Whether such use can constitute fair use directly determines the compliance costs of training data and the scope of compensation for right holders. In doctrine and judicial practice, boundaries have been drawn, respectively, by the stage of use, the nature of the act, and the purpose of the actor. The three approaches yield mutually exclusive conclusions, and their standards are incommensurable. The fundamental reason is that the basis for delimitation is drawn from industrial-policy balancing rather than from the rights structure of copyright itself. The exclusive effect of copyright extends only to the expressive interests of a work; the boundary should therefore be drawn according to whether the use effects an expressive substitution. Specifically, non-expressive uses at the input and training stages do not fall within the scope of copyright; at the output end, an expressive reproduction that is substantially similar falls within the scope of the exclusive rights; memorisation is to be assessed at the output end rather than at the training stage; and expectations of revenue in the market for licensing training corpora are not legitimate interests within the meaning of the three-step test, owing to the circularity of the argument and to the absence of any object that can be priced. Constrained by the over-expansion of the reproduction right and the closed enumeration in Article 24 of the Copyright Law, this criterion currently has no room for application under existing law. China should stipulate use as a work as a constitutive element of copyright infringement and add to the Implementing Regulations of the Copyright Law a text and data mining (TDM) exception accompanied by a machine-readable rights reservation mechanism.